[HBASE-2486] Add simple "anti-entropy" for region assignment

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 0.20.5
• Component/s: master, regionserver
• Labels: moved_from_0_20_5
• Assignee: Unassigned
• Reporter: Todd Lipcon
• Created: 25/Apr/10 20:11
• Updated: 11/Jun/22 23:32
• Resolved: 24/Dec/14 19:40

Description

We've seen a number of bugs where a region server thinks it should not be serving a region, but the master and META think it should be. I'd like to propose a very simple way of fixing this issue:

1) whenever a regionserver throws a NotServingRegionException, it also marks that region id in an RS-wide Set
2) when a region sends a heartbeat, include a message for each of these regions, MSG_REPORT_NSRE or somesuch, and then clear the set
3) when the master receives MSG_REPORT_NSRE, it does the following checks:
a) if the region is assigned elsewhere according to META, the NSRE was due to a stale client, ignore
b) if the region is in transition, ignore
c) otherwise, we have an inconsistency, and we should take some steps to resolve (eg mark the region unassigned, or exit the master if we are in "paranoid mode")

Whatever we do, we need to make sure that this is loudly logged, and causes unit tests to fail, when it's detected. This should not happen, but when it does, it would be good to recover without addtable.rb, etc.

Issue Links

• blocks: HBASE-1502 Remove need for heartbeats in HBase
• is related to: HBASE-2516 Ugly IOE when region is being closed; rather, should NSRE
• relates to: HBASE-2408 Add envelope around client<->server communication so can pass state along w/ data during interchange
• relates to: HBASE-2526 Fixup script that scans .META., report holes if any and then finds dropped regions in fs if they are present (loading them if asked)